Final_Project_ML

Overview:

Using MLAgents and Unity we are going to create a brain that when attached to a ball can solve a randomly generated maze using rays/vectors shot out to determine a path. Penalties for dead ends and reward for getting closer to end. Ultimately, I want a human player to be racing the trained agent to see who can complete the maze faster.

Machine Learning Elements:

• Deep Reinforcement learning
• Rewards and penalties
• Ray/Vector based, not necessarily based off transform position
• Both elements may be needed to prevent backtracking

Additional Required Elements:

• 3D Generated Maze
• Checkpoint asset
• Path asset
• Cube and Sphere materials/shaders

Milestone 1:

• Scene done with a static maze
• An agent that we move through the maze
• Play with rays and cameras for the trained agents
• Picking assets and deciding on maze size

Milestone 2:

• Final decision on rays and cameras
• Somewhat well trained, can get through the maze (even if slowly)

Milestone 3:

• Update maze generator to inlcude spawn and endpoints
• Using ml-agents cumulative training with the maze generator
• Randomize end marker loactions to test left, right, up, and down.

Milestone 4:

• Inlcude positive rewards for visiting new tiles
• Final train to solve as big as we can get it to
• Add player to compete with

NOTE: For the final milestone, we ran Tom's brain during the presentation.